PLACE-BASED NAVIGATION FOR AUTONOMOUS VEHICLES WITH DEEP LEARNING NEURAL NETWORKS

Accurate navigation is crucial to successfully deploying autonomous vehicles but is often limited by GPS reliance. The purpose of this thesis was to develop an image-based navigation solution that does not require GPS. This work builds on an ongoing research project, which is to develop a mobile robot that can navigate anywhere on the Naval Postgraduate School campus, inside and outside of buildings. This work focused on indoor navigation using image classification, and images were classified using a convolutional neural network (CNN). Transfer learning was used to reduce CNN training time and increase learning efficiency. The trained CNN was integrated into a waypoint loop algorithm that ran autonomously on the robot, stopping when it had correctly identified its desired location. Tests of the CNN classification showed a high success rate, but that it was also susceptible to variations in light and similar types of images. Increasing the data set will improve the classification results and allow for use in a variety of applications. This system will also benefit from improved indoor localization techniques.